<p>Recent technological developments underpinning advanced analytical methods enable safety critical entities to unravel insights from voluminous data accumulated in their repositories. A prominent disruptive technology that is gaining momentum in computational linguistics is large language models (LLMs) which hitherto have made a scant appearance in safety analytics literature on workplace safety. Therefore, this study explores LLMs’ capabilities that bear potential in safety analytics for proactive workplace safety management through two research objectives, namely: to create a theoretical synthesis of LLMs’ intrinsic capabilities; and to engender a management guidance framework for LLMs’ adoption in safety analytics. An interpretivist philosophical perspective with inductive reasoning is adopted to scrutinise LLM literature in two consecutive steps. First, the most efficacious database, search keywords and inclusion criteria are explored, resulting in the inclusion of 10,061 publications from the Scopus database. Second, an intensive study of these publications is performed through manual review of their abstracts, out of which full texts of 973 research items are examined through qualitative content analysis using NVivo. 512 nodes emerged through inductive coding, in which nodes describing LLMs’ disadvantages preponderate over the nodes on their advantages. Subsequently, a rich synthesis of LLMs’ capabilities was created in tabular format and thereafter transformed into a conceptual mapping of LLMs’ capabilities depicting: (1) core capabilities; (2) qualities occurring through their use; and (3) long-term positive and negative influences. Research findings culminate in the development of a novel theoretical framework for safety analytics establishment. This framework incorporates progressive levels of safety analytics, alongside the types of heterogeneous data (emerging in design/planning and execution/maintenance stages).</p>